All posts

WebMCP 101

The web is learning to speak agent.

WebMCP gives a website a clear way to tell an AI agent what it can do. WebMCPify helps make those capabilities easier to create, find, protect, verify, and maintain.

THE IDEA IN ONE SENTENCE

A website should offer an agent a small menu of exact actions—not a maze of pixels to guess through.

Human interface stays visible
Agent interface becomes explicit

Picture asking an agent to find a waterproof backpack, compare two options, and add your choice to the cart. Today, many agents must read the page, guess which button matters, click, and hope the website changed as expected.

WebMCP offers a cleaner route. The page can publish named tools such as search_products, add_to_cart, or check_order, together with a description and a strict input shape. The agent can understand the available action before using it, while the website’s own code keeps the visible interface in sync.

Why it is needed

Agents need handles, not guesswork.

01

Clear actions

A named tool is easier to understand than a shifting button or a hidden menu.

02

Better inputs

A schema can say exactly what is required before the agent starts an action.

03

Shared state

The website performs the action, updates its interface, and returns a structured result.

See the idea

From a request to a checked result.

Our illustrative walkthrough shows Search locating a website capability and Guard protecting the user as the action moves from choice to consent, execution, and evidence.

WebMCPify flowPLAY WHEN READY
A 48-second product walkthrough with burned-in captions and no audio.

Why this could be big

The internet becomes searchable by what it can do.

Search engines made pages findable by their words. WebMCP could make websites understandable by their actions. A travel site could expose search_routes; a bank could expose a carefully controlled download_statement; a support page could expose open_ticket.

That does not replace the human web. It adds a second, structured interface for agents. If the proposal earns broad adoption, the shift could be as meaningful as moving from pages we only read to services software can reliably act through.

Real momentum, honest status

An experiment with serious builders around it.

WebMCP is a proposed, experimental standard—not a finished universal feature. Its original proposal credits contributors from Microsoft and Google. Chrome documents the API and an origin trial. OpenAI’s WebMCP Challenge says WebMCP can be tested in ChatGPT’s in-app browser and names support from teams across Chrome, Shopify, Cloudflare, Netlify, Vercel, and Render.

That is strong ecosystem interest, not a promise that every browser, agent, or website supports WebMCP today—and it is not an endorsement of WebMCPify.

The hard part

“Callable” is not the same as “trustworthy.”

A tool can be unclear, request too much data, cross an origin boundary, trigger a consequential action, or return content an agent should not trust. Tools are also page-scoped: an agent often has to visit a page before it knows what is there. And when a website changes, a once-good capability can quietly drift.

These are infrastructure problems. They need creation discipline, useful discovery, user-owned protection, real outcome checks, and continuous monitoring—not one more optimistic success message.

What WebMCPify does

One capability. A dependable journey.

Improvisus is building WebMCPify as the capability layer around the protocol:

CreateDiscoverProtectExecuteVerifyMonitor

The future we want

Ask clearly. Act carefully. Prove what happened.

WebMCP can give agents a better way to use the web. WebMCPify is being built so that better way is also discoverable, reviewable, user-controlled, and verifiable.

Primary sources

Read the proposal yourself.

Chrome’s WebMCP guideThe API, current experiment, use cases, and limitations.WebMCP proposalThe community draft, examples, architecture, and early contributors.Chrome’s security guidanceAnnotations, consequential actions, origins, and careful exposure.OpenAI WebMCP ChallengeExamples, ecosystem participation, and ChatGPT testing information.